Models Are Getting Dumber on Purpose
Points and comments are a snapshot, not live.
AI labs trade world knowledge for reasoning skill as models shrink per parameter.
Reasoning benchmarks like AIME 2026 show scores climbing while active parameters drop: GLM-5.2 scores 99.2% with ~40B active, DeepSeek V4-Flash uses 13B. On SimpleQA, the leader Gemini 2.5 Pro scores 53%; small models hallucinate 80-82% on knowledge benchmarks. Facts take ~2 bits per parameter, while reasoning compresses better and does not go stale. The author argues retrieval and tool-use harnesses should supply facts at runtime, making small, locally-runnable reasoning models viable and reducing unfixable hallucination from weights.
What commenters are saying
Commenters split on whether the trend toward tool-augmented small models serves non-coding users. Some note most AI use is for general curiosity and text processing, not agentic harnesses. One commenter envisions pluggable knowledge bases, allowing users to compose custom models per task. Others debate whether multi-agent systems or dynamic MoE already achieve this. A correction notes SimpleQA stopped measuring in Sept 2025, pointing to SimpleQA Verified as an updated benchmark.